Stable Diffusion Models, or checkpoint models, are pre-trained Stable Diffusion weights for generating a particular style of images.
What kind of images a model generates depends on the training images. A model won’t be able to generate a cat’s image if there’s never a cat in the training data. Likewise, if you only train a model with cat images, it will only generate cats.
We will introduce what models are, some popular ones, and how to install, use, and merge them.
This is part 4 of the beginner’s guide series.
Read part 1: Absolute beginner’s guide.
Read part 2: Prompt building.
Read part 3: Inpainting.
Table of Contents
- Fine-tuned models
- What is fine-tuning?
- Why do people make Stable Diffusion models?
- How are models created?
- Popular Stable Diffusion Models
- Stable diffusion v1.4
- Stable diffusion v1.5
- Realistic Vision
- DreamShaper
- SDXL model
- Anything V3
- Best Stable Diffusion Models
- Deliberate v2
- F222
- ChilloutMix
- Protogen v2.2 (Anime)
- GhostMix
- Waifu-diffusion
- Inkpunk Diffusion
- Finding more models
- Stable Diffusion v2 models
- How to install and use a model
- CLIP Skip
- What is CLIP Skip?
- Setting CLIP Skip in AUTOMATIC1111
- Merging two models
- Example of a merged model
- Stable Diffusions model file formats
- Pruned vs Full vs EMA-only models
- Fp16 and fp32 models
- Safetensor models
- Other model types
- Summary
Fine-tuned models
What is fine-tuning?
Fine-tuning is a common technique in machine learning. It takes a model trained on a wide dataset and trains a bit more on a narrow dataset.
A fine-tuned model tends to generate images similar to those used in the training. For example, the Anything v3 model is trained with anime images. So, it generates anime-style images by default.
Why do people make Stable Diffusion models?
The Stable diffusion base model is great but is not good at everything. For example, it generates anime-style images with the keyword “anime” in the prompt. However, it could be difficult to generate images of a sub-genre of anime. Instead of tinkering with the prompt, you can use a custom model fine-tuned with images of that sub-genre.
Images below are generated using the same prompts and settings but with different models.
- Realistic Vision: Realistic photo style.
- Anything v3: Anime style.
- Dreamshaper: Realistic painting style.
Using a model is an easy way to achieve a certain style.
How are models created?
Custom checkpoint models are made with (1) additional training and (2) Dreambooth. They both start with a base model like Stable Diffusion v1.5 or XL.
Additional training is achieved by training a base model with an additional dataset you are interested in. For example, you can train the Stable Diffusion v1.5 with an additional dataset of vintage cars to bias the aesthetic of cars towards the vintage sub-genre.
Dreambooth, developed by Google, is a technique to inject custom subjects into text-to-image models. It works with as few as 3-5 custom images. You can take a few pictures of yourself and use Dreambooth to put yourself into the model. A model trained with Dreambooth requires a special keyword to condition the model.
The checkpoint model is not the only model type. We also have textual inversion (also called embedding), LoRA, LyCORIS, and hypernetwork. We will focus on the checkpoint model in this article.
Popular Stable Diffusion Models
There are thousands of fine-tuned Stable Diffusion models. The number is increasing every day. Below is a list of models that can be used for general purposes.
Stable diffusion v1.4
Released in August 2022 by Stability AI, the v1.4 model is the first publicly available Stable Diffusion model.
It is a general-purpose model, capable of producing various styles.
Most people has moved on to the v1.5 model since its release.
Stable diffusion v1.5
v1.5 was released in Oct 2022 by Runway ML, a partner of Stability AI. The model is based on v1.2 with further training.
The model page does not mention what the improvement is. It produces slightly different results compared to v1.4, but it is unclear if they are better.
Like v1.4, you can treat v1.5 as a general-purpose model. In my experience, v1.5 is a fine choice as the initial model and can be used interchangeably with v1.4.
Realistic Vision
Realistic Vision v2 is good for generating anything realistic, whether they are people, objects, or scenes.
Learn more about generating realistic people.
DreamShaper
Dreamshaper model is fine-tuned for a portrait illustration style that sits between photorealistic and computer graphics. It’s easy to use, and you will like it if you like this style.
SDXL model
SDXL model is an upgrade to the celebrated v1.5 and the forgotten v2 models.
The benefits of using the SDXL model are
- Higher native resolution – 1024 px compared to 512 px for v1.5
- Higher image quality (compared to the v1.5 base model)
- Capable of generating legible text
- It is easy to generate darker images
Anything V3
Anything V3 is a special-purpose model trained to produce high-quality anime-style images. You can use danbooru tags (like 1girl, white hair) in the text prompt.
It’s useful for casting celebrities to amine style, which can then be blended seamlessly with illustrative elements.
One drawback (at least to me) is that it produces females with disproportional body shapes. I like to tone it down with F222.
Best Stable Diffusion Models
There are thousands of Stable Diffusion models available. Many of them are special-purpose models designed to generate a particular style. Where should you start?
In addition to the ones I just mentioned, here are some of the best models I keep going back to.
Deliberate v2
Deliberate v2 is another must-have model (so many!) that renders realistic illustrations. The results can be surprisingly good. Whenever you have a good prompt, switch to this model and see what you get!
F222
F222 is trained originally for generating nudes, but people found it helpful in generating beautiful female portraits with correct body part relations. Contrary to what you might think, it’s quite good at generating aesthetically pleasing clothing.
F222 is good for portraits. It has a high tendency to generate nudes. Include wardrobe terms like “dress” and “jeans” in the prompt.
Find more realistic photo-style models in this post.
ChilloutMix
ChilloutMix is a special model for generating photo-quality Asian females. It is like the Asian counterpart of F222. Use with Korean embedding ulzzang-6500-v1 to generate girls like k-pop.
Like F222, it generates nudes sometimes. Suppress with wardrobe terms like “dress” and “jeans” in the prompt, and “nude” in the negative prompt.
Protogen v2.2 (Anime)
Protogen v2.2 is classy. It generates illustration and anime-style images with good taste.
GhostMix
GhostMix is trained with Ghost in the Shell style, a classic anime in the 90s. You will find it useful for generating cyborgs and robots.
Waifu-diffusion
Waifu Diffusion is a Japanese anime style.
Inkpunk Diffusion
Inkpunk Diffusion is a Dreambooth-trained model with a very distinct illustration style.
Use keyword: nvinkpunk
Finding more models
Civitai is the go-to place for downloading models.
Huggingface is another good source though the interface is not designed for Stable Diffusion models.
Stable Diffusion v2 models
Stable Diffusion v2 are two official Stable Diffusion models. The main change in v2 models are
- In addition to 512×512 pixels, a higher resolution version of 768×768 pixels is available.
- You can no longer generate explicit content because p*rnographic materials were removed from training.
You may assume that everyone has moved on to using the v2 models from v1.5. However, the Stable Diffusion community found that the images looked worse in the 2.0 model. People also have difficulty using powerful keywords like celebrity and artist names.
The 2.1 model has partially addressed these issues. The images look better out of the box. It’s easier to generate artistic style.
Now, most people don’t use v2 models. We use v1 and SDXL. But if you want to try out v2 models, check out these tips to avoid common frustrations.
How to install and use a model
These instructions are for v1 and SDXL models.
To install a model in AUTOMATIC1111 GUI, download and place the checkpoint model file in the following folder
stable-diffusion-webui/models/Stable-diffusion/Press the reload button next to the checkpoint dropbox on top left.
You should see the checkpoint file you just put in available for selection. Select the new checkpoint file to use the model.
Alternatively, select the Checkpoints tab on the txt2img or img2img page and choose a model.
If you are new to AUTOMATIC1111 GUI, some models are preloaded in the Colab notebook included in theQuick Start Guide.
See the instructions for v2.0 and v2.1.
See the SDXL article for using the SDXL model.
CLIP Skip
Some models recommend a different Clip Skip setting. You should follow this setting to get the intended style.
What is CLIP Skip?
CLIP Skip is a feature that skips a number of last layers in the CLIP text embedding network during image generation in Stable Diffusion. CLIP is the language model used in Stable Diffusion v1.5 models. It converts text tokens in the prompt into embeddings. It is a deep neural network model that contains many layers. CLIP Skip refers to how many of the last layers to skip.
In AUTOMATIC1111 and many Stable Diffusion software, CLIP Skip of 1 does not skip any layers. CLIP Skip of 2 skips the last layer, and so on.
Why do you want to skip some CLIP layers? A neural network summarizes information as it passes through layers. The earlier the layer is, the richer the information it contains.
Skipping CLIP layers can have a dramatic effect on images. Many anime models are trained with CLIP Skip of 2. See the examples below with different CLIP skips but the same prompt and seed.
Setting CLIP Skip in AUTOMATIC1111
You can set the CLIP Skip on the Settings Page > Stable Diffusion > Clip Skip. Adjust the value and click Apply Settings.
But if you need to change CLIP Skip regularly, a better way is to add it to the Quick Settings. Go to the Settings page > User Interface > Quicksettings list. Add CLIP_stop_at_last_layer. Click Apply Settings and Reload UI.
The clip skip slider should appear at the top of AUTOMATIC1111.
Merging two models
To merge two models using AUTOMATIC1111 GUI, go to the Checkpoint Merger tab and select the two models you want to merge in Primary model (A) and Secondary model (B).
Adjust the multiplier (M) to adjust the relative weight of the two models. Setting it to 0.5 would merge the two models with equal importance.
After pressing Run, the new merged model will be available for use.
Example of a merged model
Here are sample images from merging F222 and Anything V3 with equal weight (0.5):
The merged model sits between the realistic F222 and the anime Anything V3 styles. It is a very good model for generating illustration art with human figures.
Stable Diffusions model file formats
On a model download page, you may see several model file formats.
- Pruned
- Full
- EMA-only
- FP16
- FP32
- .pt
- .safetensor
This is confusing! Which one should you download?
Pruned vs Full vs EMA-only models
Some Stable Diffusion checkpoint models consist of two sets of weights: (1) The weights after the last training step and (2) the average weights over the last few training steps, called EMA (exponential moving average).
You can download the EMA-only model if you are only interested in using the model. These are the weights you use when you use the model. They are sometimes called pruned models.
You will only need the full model (i.e. A checkpoint file consisting of two sets of weights) if you want to fine-tune the model with additional training.
So, download the pruned or EMA-only model if you want to use it to generate images. This saves you some disk space. Trust me, your hard drive will fill up very soon!
Fp16 and fp32 models
FP stands for floating point. It is a computer’s way of storing decimal numbers. Here the decimal numbers are the model weights. FP16 takes 16 bits per number and is called half precision. FP32 takes 32 bits and is called full precision.
The training data for deep learning models (such as Stable Diffusion) is pretty noisy. You rarely need a full-precision model. The extra precision just stores noise!
So, download the FP16 models if available. They are about half as big. This saves you a few GB!
Safetensor models
The original pytorch model format is .pt. The downside of this format is that it is not secure. Someone can pack malicious codes in it. The code can run on your machine when you use the model.
Safetensors is an improved version of the PT model format. It does the same thing as storing the weights but will not execute any codes.
So, download the safetensors version whenever it is available. If not, download the PT files from a trustworthy source.
Other model types
Four main types of files can be called “models”. Let’s clarify them so you know what people are talking about.
- Checkpoint models are the real Stable Diffusion models. They contain all you need to generate an image. No additional files are required. They are large, typically 2 – 7 GB. They are the subject of this article.
- Textual inversions (also called embeddings) are small files defining new keywords to generate new objects or styles. They are small, typically 10 – 100 KB. You must use them with a checkpoint model.
- LoRA models are small patch files to checkpoint models for modifying styles. They are typically 10-200 MB. You must use them with a checkpoint model.
- Hypernetworks are additional network modules added to checkpoint models. They are typically 5 – 300 MB. You must use them with a checkpoint model.
Summary
I have introduced Stable Diffusion models, how they are made, a few common ones, and how to merge them. Using models can make your life easier when you have a specific style in mind.
Check out the Stable Diffusion Course for a step-by-step guided course.
This is part 4 of the beginner’s guide series.
Read part 1: Absolute beginner’s guide.
Read part 2: Prompt building.
Read part 3: Inpainting.
